Objective: The primary aim of this study was to use metagenomic next-generation sequencing (mNGS) data to identify coronavirus 2019 (COVID-19)-related biomarker genes and to construct a machine learning model that could successfully differentiate patients with COVID-19 from healthy controls.
Materials and Methods: The mNGS dataset used in the study demonstrated expression of 15,979 genes in the upper air-way in 234 patients who were COVID-19 negative and COVID-19 positive. The Boruta method was used to select qualitative biomarker genes associated with COVID-19. Random forest (RF), gradient boosting tree (GBT), and multi-layer perceptron (MLP) models were used to predict COVID-19 based on the selected biomarker genes.
Results: The MLP (0.936) model outperformed the GBT (0.851), and RF (0.809) models in predicting COVID-19. The three most important biomarker candidate genes associated with COVID-19 were IFI27, TPTI, and FAM83A.
Conclusion: The proposed model (MLP) was able to predict COVID-19 successfully. The results showed that the generated model and selected biomarker candidate genes can be used as diagnostic models for clinical testing or potential therapeutic targets and vaccine design.